4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2013, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/obdclass/obd_mount_server.c
34 * Server mount routines
36 * Author: Nathan Rutman <nathan@clusterfs.com>
40 #define DEBUG_SUBSYSTEM S_CLASS
41 #define D_MOUNT (D_SUPER | D_CONFIG /* | D_WARNING */)
42 #define PRINT_CMD CDEBUG
43 #define PRINT_MASK (D_SUPER | D_CONFIG)
45 #include <linux/types.h>
46 #ifdef HAVE_LINUX_SELINUX_IS_ENABLED
47 #include <linux/selinux.h>
49 #include <linux/statfs.h>
50 #include <linux/version.h>
51 #include <linux/delay.h>
53 #include <llog_swab.h>
54 #include <lustre_disk.h>
55 #include <uapi/linux/lustre/lustre_ioctl.h>
56 #include <lustre_log.h>
57 #include <uapi/linux/lustre/lustre_param.h>
59 #include <obd_class.h>
61 /*********** mount lookup *********/
63 static DEFINE_MUTEX(lustre_mount_info_lock);
64 static LIST_HEAD(server_mount_info_list);
66 static struct lustre_mount_info *server_find_mount(const char *name)
68 struct list_head *tmp;
69 struct lustre_mount_info *lmi;
72 list_for_each(tmp, &server_mount_info_list) {
73 lmi = list_entry(tmp, struct lustre_mount_info,
75 if (strcmp(name, lmi->lmi_name) == 0)
81 /* we must register an obd for a mount before we call the setup routine.
82 *_setup will call lustre_get_mount to get the mnt struct
83 by obd_name, since we can't pass the pointer to setup. */
84 static int server_register_mount(const char *name, struct super_block *sb)
86 struct lustre_mount_info *lmi;
92 OBD_ALLOC(lmi, sizeof(*lmi));
95 OBD_ALLOC(name_cp, strlen(name) + 1);
97 OBD_FREE(lmi, sizeof(*lmi));
100 strcpy(name_cp, name);
102 mutex_lock(&lustre_mount_info_lock);
104 if (server_find_mount(name)) {
105 mutex_unlock(&lustre_mount_info_lock);
106 OBD_FREE(lmi, sizeof(*lmi));
107 OBD_FREE(name_cp, strlen(name) + 1);
108 CERROR("Already registered %s\n", name);
111 lmi->lmi_name = name_cp;
113 list_add(&lmi->lmi_list_chain, &server_mount_info_list);
115 mutex_unlock(&lustre_mount_info_lock);
117 CDEBUG(D_MOUNT, "register mount %p from %s\n", sb, name);
122 /* when an obd no longer needs a mount */
123 static int server_deregister_mount(const char *name)
125 struct lustre_mount_info *lmi;
128 mutex_lock(&lustre_mount_info_lock);
129 lmi = server_find_mount(name);
131 mutex_unlock(&lustre_mount_info_lock);
132 CERROR("%s not registered\n", name);
136 CDEBUG(D_MOUNT, "deregister mount %p from %s\n", lmi->lmi_sb, name);
138 OBD_FREE(lmi->lmi_name, strlen(lmi->lmi_name) + 1);
139 list_del(&lmi->lmi_list_chain);
140 OBD_FREE(lmi, sizeof(*lmi));
141 mutex_unlock(&lustre_mount_info_lock);
146 /* obd's look up a registered mount using their obdname. This is just
147 for initial obd setup to find the mount struct. It should not be
148 called every time you want to mntget. */
149 struct lustre_mount_info *server_get_mount(const char *name)
151 struct lustre_mount_info *lmi;
152 struct lustre_sb_info *lsi;
155 mutex_lock(&lustre_mount_info_lock);
156 lmi = server_find_mount(name);
157 mutex_unlock(&lustre_mount_info_lock);
159 CERROR("Can't find mount for %s\n", name);
162 lsi = s2lsi(lmi->lmi_sb);
164 atomic_inc(&lsi->lsi_mounts);
166 CDEBUG(D_MOUNT, "get mount %p from %s, refs=%d\n", lmi->lmi_sb,
167 name, atomic_read(&lsi->lsi_mounts));
171 EXPORT_SYMBOL(server_get_mount);
174 * server_put_mount: to be called from obd_cleanup methods
176 * @dereg_mnt: 0 or 1 depending on whether the mount is to be deregistered or
179 * The caller decides whether server_deregister_mount() needs to be called or
180 * not. Calling of server_deregister_mount() does not depend on refcounting on
181 * lsi because we could have say the mgs and mds on the same node and we
182 * unmount the mds, then the ref on the lsi would still be non-zero but we
183 * would still want to deregister the mds mount.
185 int server_put_mount(const char *name, bool dereg_mnt)
187 struct lustre_mount_info *lmi;
188 struct lustre_sb_info *lsi;
191 mutex_lock(&lustre_mount_info_lock);
192 lmi = server_find_mount(name);
193 mutex_unlock(&lustre_mount_info_lock);
195 CERROR("Can't find mount for %s\n", name);
198 lsi = s2lsi(lmi->lmi_sb);
200 CDEBUG(D_MOUNT, "put mount %p from %s, refs=%d\n",
201 lmi->lmi_sb, name, atomic_read(&lsi->lsi_mounts));
203 if (lustre_put_lsi(lmi->lmi_sb))
204 CDEBUG(D_MOUNT, "Last put of mount %p from %s\n",
208 /* this obd should never need the mount again */
209 server_deregister_mount(name);
213 EXPORT_SYMBOL(server_put_mount);
215 /* Set up a MGS to serve startup logs */
216 static int server_start_mgs(struct super_block *sb)
218 struct lustre_sb_info *lsi = s2lsi(sb);
219 struct lustre_mount_info *lmi;
223 /* It is impossible to have more than 1 MGS per node, since
224 MGC wouldn't know which to connect to */
225 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
227 lsi = s2lsi(lmi->lmi_sb);
228 LCONSOLE_ERROR_MSG(0x15d, "The MGS service was already started"
233 CDEBUG(D_CONFIG, "Start MGS service %s\n", LUSTRE_MGS_OBDNAME);
235 rc = server_register_mount(LUSTRE_MGS_OBDNAME, sb);
238 rc = lustre_start_simple(LUSTRE_MGS_OBDNAME, LUSTRE_MGS_NAME,
239 LUSTRE_MGS_OBDNAME, NULL, NULL,
240 lsi->lsi_osd_obdname, NULL);
241 /* server_deregister_mount() is not called previously, for lsi
242 * and other stuff can't be freed cleanly when mgs calls
243 * server_put_mount() in error handling case (see b=17758),
244 * this problem is caused by a bug in mgs_init0, which forgot
245 * calling server_put_mount in error case. */
248 server_deregister_mount(LUSTRE_MGS_OBDNAME);
252 LCONSOLE_ERROR_MSG(0x15e, "Failed to start MGS '%s' (%d). "
253 "Is the 'mgs' module loaded?\n",
254 LUSTRE_MGS_OBDNAME, rc);
258 static int server_stop_mgs(struct super_block *sb)
260 struct obd_device *obd;
262 struct lustre_mount_info *lmi;
265 /* Do not stop MGS if this device is not the running MGT */
266 lmi = server_find_mount(LUSTRE_MGS_OBDNAME);
267 if (lmi != NULL && lmi->lmi_sb != sb)
270 CDEBUG(D_MOUNT, "Stop MGS service %s\n", LUSTRE_MGS_OBDNAME);
272 /* There better be only one MGS */
273 obd = class_name2obd(LUSTRE_MGS_OBDNAME);
275 CDEBUG(D_CONFIG, "mgs %s not running\n", LUSTRE_MGS_OBDNAME);
279 /* The MGS should always stop when we say so */
281 rc = class_manual_cleanup(obd);
285 /* Since there's only one mgc per node, we have to change it's fs to get
286 access to the right disk. */
287 static int server_mgc_set_fs(const struct lu_env *env,
288 struct obd_device *mgc, struct super_block *sb)
290 struct lustre_sb_info *lsi = s2lsi(sb);
294 CDEBUG(D_MOUNT, "Set mgc disk for %s\n", lsi->lsi_lmd->lmd_dev);
296 /* cl_mgc_sem in mgc insures we sleep if the mgc_fs is busy */
297 rc = obd_set_info_async(env, mgc->obd_self_export,
298 sizeof(KEY_SET_FS), KEY_SET_FS,
299 sizeof(*sb), sb, NULL);
301 CERROR("can't set_fs %d\n", rc);
306 static int server_mgc_clear_fs(const struct lu_env *env,
307 struct obd_device *mgc)
312 CDEBUG(D_MOUNT, "Unassign mgc disk\n");
314 rc = obd_set_info_async(env, mgc->obd_self_export,
315 sizeof(KEY_CLEAR_FS), KEY_CLEAR_FS,
320 static inline bool is_mdc_device(const char *devname)
324 ptr = strrchr(devname, '-');
325 return ptr != NULL && strcmp(ptr, "-mdc") == 0;
328 static inline bool tgt_is_mdt(const char *tgtname, __u32 *idx)
332 type = server_name2index(tgtname, idx, NULL);
334 return type == LDD_F_SV_TYPE_MDT;
338 * Convert OST/MDT name(fsname-{MDT,OST}xxxx) to a lwp name with the @idx:yyyy
339 * (fsname-MDTyyyy-lwp-{MDT,OST}xxxx)
341 int tgt_name2lwp_name(const char *tgt_name, char *lwp_name, int len, __u32 idx)
348 OBD_ALLOC(fsname, MTI_NAME_MAXLEN);
352 rc = server_name2fsname(tgt_name, fsname, &tgt);
354 CERROR("%s: failed to get fsname from tgt_name: rc = %d\n",
359 if (*tgt != '-' && *tgt != ':') {
360 CERROR("%s: invalid tgt_name name!\n", tgt_name);
361 GOTO(cleanup, rc = -EINVAL);
365 if (strncmp(tgt, "OST", 3) != 0 && strncmp(tgt, "MDT", 3) != 0) {
366 CERROR("%s is not an OST or MDT target!\n", tgt_name);
367 GOTO(cleanup, rc = -EINVAL);
369 snprintf(lwp_name, len, "%s-MDT%04x-%s-%s",
370 fsname, idx, LUSTRE_LWP_NAME, tgt);
372 GOTO(cleanup, rc = 0);
376 OBD_FREE(fsname, MTI_NAME_MAXLEN);
380 EXPORT_SYMBOL(tgt_name2lwp_name);
382 static LIST_HEAD(lwp_register_list);
383 static DEFINE_SPINLOCK(lwp_register_list_lock);
385 static void lustre_put_lwp_item(struct lwp_register_item *lri)
387 if (atomic_dec_and_test(&lri->lri_ref)) {
388 LASSERT(list_empty(&lri->lri_list));
390 if (*lri->lri_exp != NULL)
391 class_export_put(*lri->lri_exp);
396 int lustre_register_lwp_item(const char *lwpname, struct obd_export **exp,
397 register_lwp_cb cb_func, void *cb_data)
399 struct obd_device *lwp;
400 struct lwp_register_item *lri;
404 LASSERTF(strlen(lwpname) < MTI_NAME_MAXLEN, "lwpname is too long %s\n",
406 LASSERT(exp != NULL && *exp == NULL);
412 lwp = class_name2obd(lwpname);
413 if (lwp != NULL && lwp->obd_set_up == 1) {
414 struct obd_uuid *uuid;
421 memcpy(uuid->uuid, lwpname, strlen(lwpname));
422 *exp = cfs_hash_lookup(lwp->obd_uuid_hash, uuid);
426 memcpy(lri->lri_name, lwpname, strlen(lwpname));
428 lri->lri_cb_func = cb_func;
429 lri->lri_cb_data = cb_data;
430 INIT_LIST_HEAD(&lri->lri_list);
432 * Initialize the lri_ref at 2, one will be released before
433 * current function returned via lustre_put_lwp_item(), the
434 * other will be released in lustre_deregister_lwp_item().
436 atomic_set(&lri->lri_ref, 2);
438 spin_lock(&lwp_register_list_lock);
439 list_add(&lri->lri_list, &lwp_register_list);
442 spin_unlock(&lwp_register_list_lock);
444 if (cb && cb_func != NULL)
446 lustre_put_lwp_item(lri);
450 EXPORT_SYMBOL(lustre_register_lwp_item);
452 void lustre_deregister_lwp_item(struct obd_export **exp)
454 struct lwp_register_item *lri;
455 bool removed = false;
458 spin_lock(&lwp_register_list_lock);
459 list_for_each_entry(lri, &lwp_register_list, lri_list) {
460 if (exp == lri->lri_exp) {
461 list_del_init(&lri->lri_list);
466 spin_unlock(&lwp_register_list_lock);
471 /* See lustre_notify_lwp_list(), in some extreme race conditions,
472 * the notify callback could be still on the fly, we need to wait
473 * for the callback done before moving on to free the data used
475 while (atomic_read(&lri->lri_ref) > 1) {
476 CDEBUG(D_MOUNT, "lri reference count %u, repeat: %d\n",
477 atomic_read(&lri->lri_ref), repeat);
479 set_current_state(TASK_INTERRUPTIBLE);
480 schedule_timeout(cfs_time_seconds(1));
482 lustre_put_lwp_item(lri);
484 EXPORT_SYMBOL(lustre_deregister_lwp_item);
486 struct obd_export *lustre_find_lwp_by_index(const char *dev, __u32 idx)
488 struct lustre_mount_info *lmi;
489 struct lustre_sb_info *lsi;
490 struct obd_device *lwp;
491 struct obd_export *exp = NULL;
496 lmi = server_get_mount(dev);
500 lsi = s2lsi(lmi->lmi_sb);
501 rc = server_name2fsname(lsi->lsi_svname, fsname, NULL);
503 CERROR("%s: failed to get fsname: rc = %d\n",
504 lsi->lsi_svname, rc);
508 snprintf(lwp_name, sizeof(lwp_name), "%s-MDT%04x", fsname, idx);
509 mutex_lock(&lsi->lsi_lwp_mutex);
510 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
511 char *ptr = strstr(lwp->obd_name, lwp_name);
513 if (ptr != NULL && lwp->obd_lwp_export != NULL) {
514 exp = class_export_get(lwp->obd_lwp_export);
518 mutex_unlock(&lsi->lsi_lwp_mutex);
521 server_put_mount(dev, false);
525 EXPORT_SYMBOL(lustre_find_lwp_by_index);
527 void lustre_notify_lwp_list(struct obd_export *exp)
529 struct lwp_register_item *lri;
530 LASSERT(exp != NULL);
533 spin_lock(&lwp_register_list_lock);
534 list_for_each_entry(lri, &lwp_register_list, lri_list) {
535 if (strcmp(exp->exp_obd->obd_name, lri->lri_name))
537 if (*lri->lri_exp != NULL)
539 *lri->lri_exp = class_export_get(exp);
540 if (lri->lri_cb_func == NULL)
542 atomic_inc(&lri->lri_ref);
543 spin_unlock(&lwp_register_list_lock);
545 lri->lri_cb_func(lri->lri_cb_data);
546 lustre_put_lwp_item(lri);
548 /* Others may have changed the list after we unlock, we have
549 * to rescan the list from the beginning. Usually, the list
550 * 'lwp_register_list' is very short, and there is 'guard'
551 * lri::lri_exp that will prevent the callback to be done
552 * repeatedly. So rescanning the list has no problem. */
555 spin_unlock(&lwp_register_list_lock);
557 EXPORT_SYMBOL(lustre_notify_lwp_list);
559 static int lustre_lwp_connect(struct obd_device *lwp, bool is_mdt)
562 struct lu_context session_ctx;
563 struct obd_export *exp;
564 struct obd_uuid *uuid = NULL;
565 struct obd_connect_data *data = NULL;
569 /* log has been fully processed, let clients connect */
570 rc = lu_env_init(&env, lwp->obd_lu_dev->ld_type->ldt_ctx_tags);
574 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
575 session_ctx.lc_thread = NULL;
576 lu_context_enter(&session_ctx);
577 env.le_ses = &session_ctx;
581 GOTO(out, rc = -ENOMEM);
583 data->ocd_connect_flags = OBD_CONNECT_VERSION | OBD_CONNECT_INDEX;
584 data->ocd_version = LUSTRE_VERSION_CODE;
585 data->ocd_connect_flags |= OBD_CONNECT_FID | OBD_CONNECT_AT |
586 OBD_CONNECT_LRU_RESIZE | OBD_CONNECT_FULL20 |
587 OBD_CONNECT_LVB_TYPE | OBD_CONNECT_LIGHTWEIGHT |
588 OBD_CONNECT_LFSCK | OBD_CONNECT_BULK_MBITS;
591 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
595 GOTO(out, rc = -ENOMEM);
597 if (strlen(lwp->obd_name) > sizeof(uuid->uuid)) {
598 CERROR("%s: Too long lwp name %s, max_size is %d\n",
599 lwp->obd_name, lwp->obd_name, (int)sizeof(uuid->uuid));
600 GOTO(out, rc = -EINVAL);
603 /* Use lwp name as the uuid, so we find the export by lwp name later */
604 memcpy(uuid->uuid, lwp->obd_name, strlen(lwp->obd_name));
605 rc = obd_connect(&env, &exp, lwp, uuid, data, NULL);
607 CERROR("%s: connect failed: rc = %d\n", lwp->obd_name, rc);
609 if (unlikely(lwp->obd_lwp_export != NULL))
610 class_export_put(lwp->obd_lwp_export);
611 lwp->obd_lwp_export = class_export_get(exp);
623 lu_context_exit(&session_ctx);
624 lu_context_fini(&session_ctx);
630 * lwp is used by slaves (Non-MDT0 targets) to manage the connection to MDT0,
631 * or from the OSTx to MDTy.
633 static int lustre_lwp_setup(struct lustre_cfg *lcfg, struct lustre_sb_info *lsi,
636 struct obd_device *obd;
637 char *lwpname = NULL;
638 char *lwpuuid = NULL;
642 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
645 CERROR("%s: Can't add uuid: rc =%d\n", lsi->lsi_svname, rc);
649 OBD_ALLOC(lwpname, MTI_NAME_MAXLEN);
651 GOTO(out, rc = -ENOMEM);
653 rc = tgt_name2lwp_name(lsi->lsi_svname, lwpname, MTI_NAME_MAXLEN, idx);
655 CERROR("%s: failed to generate lwp name: rc = %d\n",
656 lsi->lsi_svname, rc);
660 OBD_ALLOC(lwpuuid, MTI_NAME_MAXLEN);
662 GOTO(out, rc = -ENOMEM);
664 sprintf(lwpuuid, "%s_UUID", lwpname);
665 rc = lustre_start_simple(lwpname, LUSTRE_LWP_NAME,
666 lwpuuid, lustre_cfg_string(lcfg, 1),
669 CERROR("%s: setup up failed: rc %d\n", lwpname, rc);
673 obd = class_name2obd(lwpname);
674 LASSERT(obd != NULL);
676 rc = lustre_lwp_connect(obd, strstr(lsi->lsi_svname, "-MDT") != NULL);
678 obd->u.cli.cl_max_mds_easize = MAX_MD_SIZE;
679 mutex_lock(&lsi->lsi_lwp_mutex);
680 list_add_tail(&obd->obd_lwp_list, &lsi->lsi_lwp_list);
681 mutex_unlock(&lsi->lsi_lwp_mutex);
683 CERROR("%s: connect failed: rc = %d\n", lwpname, rc);
690 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
692 OBD_FREE(lwpuuid, MTI_NAME_MAXLEN);
697 /* the caller is responsible for memory free */
698 static struct obd_device *lustre_find_lwp(struct lustre_sb_info *lsi,
699 char **lwpname, __u32 idx)
701 struct obd_device *lwp;
705 LASSERT(lwpname != NULL);
706 LASSERT(IS_OST(lsi) || IS_MDT(lsi));
708 OBD_ALLOC(*lwpname, MTI_NAME_MAXLEN);
709 if (*lwpname == NULL)
710 RETURN(ERR_PTR(-ENOMEM));
712 rc = tgt_name2lwp_name(lsi->lsi_svname, *lwpname, MTI_NAME_MAXLEN, idx);
714 CERROR("%s: failed to generate lwp name: rc = %d\n",
715 lsi->lsi_svname, rc);
716 GOTO(out, rc = -EINVAL);
719 lwp = class_name2obd(*lwpname);
723 if (*lwpname != NULL) {
724 OBD_FREE(*lwpname, MTI_NAME_MAXLEN);
730 RETURN(lwp != NULL ? lwp : ERR_PTR(-ENOENT));
733 static int lustre_lwp_add_conn(struct lustre_cfg *cfg,
734 struct lustre_sb_info *lsi, __u32 idx)
736 struct lustre_cfg_bufs *bufs = NULL;
737 struct lustre_cfg *lcfg = NULL;
738 char *lwpname = NULL;
739 struct obd_device *lwp;
743 lwp = lustre_find_lwp(lsi, &lwpname, idx);
745 CERROR("%s: can't find lwp device.\n", lsi->lsi_svname);
746 GOTO(out, rc = PTR_ERR(lwp));
748 LASSERT(lwpname != NULL);
752 GOTO(out, rc = -ENOMEM);
754 lustre_cfg_bufs_reset(bufs, lwpname);
755 lustre_cfg_bufs_set_string(bufs, 1,
756 lustre_cfg_string(cfg, 1));
758 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount, bufs->lcfg_buflen));
760 GOTO(out_cfg, rc = -ENOMEM);
761 lustre_cfg_init(lcfg, LCFG_ADD_CONN, bufs);
763 rc = class_add_conn(lwp, lcfg);
765 CERROR("%s: can't add conn: rc = %d\n", lwpname, rc);
768 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
769 lcfg->lcfg_buflens));
775 OBD_FREE(lwpname, MTI_NAME_MAXLEN);
780 * Retrieve MDT nids from the client log, then start the lwp device.
781 * there are only two scenarios which would include mdt nid.
783 * marker 5 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxx-
784 * add_uuid nid=192.168.122.162@tcp(0x20000c0a87aa2) 0: 1:192.168.122.162@tcp
785 * attach 0:lustre-MDTyyyy-mdc 1:mdc 2:lustre-clilmv_UUID
786 * setup 0:lustre-MDTyyyy-mdc 1:lustre-MDTyyyy_UUID 2:192.168.122.162@tcp
787 * add_uuid nid=192.168.172.1@tcp(0x20000c0a8ac01) 0: 1:192.168.172.1@tcp
788 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.172.1@tcp
789 * modify_mdc_tgts add 0:lustre-clilmv 1:lustre-MDTyyyy_UUID xxxx
790 * marker 5 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add mdc' xxxx-
792 * marker 7 (flags=0x01, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
793 * add_uuid nid=192.168.122.2@tcp(0x20000c0a87a02) 0: 1:192.168.122.2@tcp
794 * add_conn 0:lustre-MDTyyyy-mdc 1:192.168.122.2@tcp
795 * marker 7 (flags=0x02, v2.1.54.0) lustre-MDTyyyy 'add failnid' xxxx-
797 static int client_lwp_config_process(const struct lu_env *env,
798 struct llog_handle *handle,
799 struct llog_rec_hdr *rec, void *data)
801 struct config_llog_instance *cfg = data;
802 int cfg_len = rec->lrh_len;
803 char *cfg_buf = (char *) (rec + 1);
804 struct lustre_cfg *lcfg = NULL;
805 struct lustre_sb_info *lsi;
806 int rc = 0, swab = 0;
809 if (rec->lrh_type != OBD_CFG_REC) {
810 CERROR("Unknown llog record type %#x encountered\n",
815 if (cfg->cfg_sb == NULL)
816 GOTO(out, rc = -EINVAL);
817 lsi = s2lsi(cfg->cfg_sb);
819 lcfg = (struct lustre_cfg *)cfg_buf;
820 if (lcfg->lcfg_version == __swab32(LUSTRE_CFG_VERSION)) {
821 lustre_swab_lustre_cfg(lcfg);
825 rc = lustre_cfg_sanity_check(cfg_buf, cfg_len);
829 switch (lcfg->lcfg_command) {
831 struct cfg_marker *marker = lustre_cfg_buf(lcfg, 1);
833 lustre_swab_cfg_marker(marker, swab,
834 LUSTRE_CFG_BUFLEN(lcfg, 1));
835 if (marker->cm_flags & CM_SKIP ||
836 marker->cm_flags & CM_EXCLUDE)
839 if (!tgt_is_mdt(marker->cm_tgtname, &cfg->cfg_lwp_idx))
842 if (IS_MDT(lsi) && cfg->cfg_lwp_idx != 0)
845 if (!strncmp(marker->cm_comment, "add mdc", 7) ||
846 !strncmp(marker->cm_comment, "add failnid", 11)) {
847 if (marker->cm_flags & CM_START) {
848 cfg->cfg_flags = CFG_F_MARKER;
849 /* This hack is to differentiate the
850 * ADD_UUID is come from "add mdc" record
851 * or from "add failnid" record. */
852 if (!strncmp(marker->cm_comment,
854 cfg->cfg_flags |= CFG_F_SKIP;
855 } else if (marker->cm_flags & CM_END) {
861 case LCFG_ADD_UUID: {
862 if (cfg->cfg_flags == CFG_F_MARKER) {
863 rc = lustre_lwp_setup(lcfg, lsi, cfg->cfg_lwp_idx);
864 /* XXX: process only the first nid as
865 * we don't need another instance of lwp */
866 cfg->cfg_flags |= CFG_F_SKIP;
867 } else if (cfg->cfg_flags == (CFG_F_MARKER | CFG_F_SKIP)) {
868 rc = class_add_uuid(lustre_cfg_string(lcfg, 1),
871 CERROR("%s: Fail to add uuid, rc:%d\n",
872 lsi->lsi_svname, rc);
876 case LCFG_ADD_CONN: {
877 char *devname = lustre_cfg_string(lcfg, 0);
881 if (!is_mdc_device(devname))
884 ptr = strrchr(devname, '-');
889 if (!tgt_is_mdt(devname, &idx)) {
895 if (IS_MDT(lsi) && idx != 0)
898 rc = lustre_lwp_add_conn(lcfg, lsi, idx);
908 static int lustre_disconnect_lwp(struct super_block *sb)
910 struct lustre_sb_info *lsi = s2lsi(sb);
911 struct obd_device *lwp;
912 char *logname = NULL;
913 struct lustre_cfg_bufs *bufs = NULL;
914 struct config_llog_instance *cfg = NULL;
919 if (likely(lsi->lsi_lwp_started)) {
920 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
924 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
926 CERROR("%s: failed to get fsname from svname: "
927 "rc = %d\n", lsi->lsi_svname, rc);
928 GOTO(out, rc = -EINVAL);
931 strcat(logname, "-client");
934 GOTO(out, rc = -ENOMEM);
937 cfg->cfg_instance = ll_get_cfg_instance(sb);
938 rc = lustre_end_log(sb, logname, cfg);
939 if (rc != 0 && rc != -ENOENT)
942 lsi->lsi_lwp_started = 0;
947 GOTO(out, rc = -ENOMEM);
949 mutex_lock(&lsi->lsi_lwp_mutex);
950 list_for_each_entry(lwp, &lsi->lsi_lwp_list, obd_lwp_list) {
951 struct lustre_cfg *lcfg;
953 if (likely(lwp->obd_lwp_export != NULL)) {
954 class_export_put(lwp->obd_lwp_export);
955 lwp->obd_lwp_export = NULL;
958 lustre_cfg_bufs_reset(bufs, lwp->obd_name);
959 lustre_cfg_bufs_set_string(bufs, 1, NULL);
960 OBD_ALLOC(lcfg, lustre_cfg_len(bufs->lcfg_bufcount,
966 lustre_cfg_init(lcfg, LCFG_CLEANUP, bufs);
968 /* Disconnect import first. NULL is passed for the '@env',
969 * since it will not be used. */
970 rc = lwp->obd_lu_dev->ld_ops->ldo_process_config(NULL,
971 lwp->obd_lu_dev, lcfg);
972 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
973 lcfg->lcfg_buflens));
974 if (rc != 0 && rc != -ETIMEDOUT) {
975 CERROR("%s: fail to disconnect LWP: rc = %d\n",
980 mutex_unlock(&lsi->lsi_lwp_mutex);
990 OBD_FREE(logname, MTI_NAME_MAXLEN);
992 return rc1 != 0 ? rc1 : rc;
996 * Stop the lwp for an OST/MDT target.
998 static int lustre_stop_lwp(struct super_block *sb)
1000 struct lustre_sb_info *lsi = s2lsi(sb);
1001 struct obd_device *lwp;
1006 mutex_lock(&lsi->lsi_lwp_mutex);
1007 while (!list_empty(&lsi->lsi_lwp_list)) {
1008 lwp = list_entry(lsi->lsi_lwp_list.next, struct obd_device,
1010 list_del_init(&lwp->obd_lwp_list);
1012 mutex_unlock(&lsi->lsi_lwp_mutex);
1014 rc = class_manual_cleanup(lwp);
1016 CERROR("%s: fail to stop LWP: rc = %d\n",
1020 mutex_lock(&lsi->lsi_lwp_mutex);
1022 mutex_unlock(&lsi->lsi_lwp_mutex);
1024 RETURN(rc1 != 0 ? rc1 : rc);
1028 * Start the lwp(fsname-MDTyyyy-lwp-{MDT,OST}xxxx) for a MDT/OST or MDT target.
1030 static int lustre_start_lwp(struct super_block *sb)
1032 struct lustre_sb_info *lsi = s2lsi(sb);
1033 struct config_llog_instance *cfg = NULL;
1038 if (unlikely(lsi->lsi_lwp_started))
1041 OBD_ALLOC(logname, MTI_NAME_MAXLEN);
1042 if (logname == NULL)
1045 rc = server_name2fsname(lsi->lsi_svname, logname, NULL);
1047 CERROR("%s: failed to get fsname from svname: rc = %d\n",
1048 lsi->lsi_svname, rc);
1049 GOTO(out, rc = -EINVAL);
1052 strcat(logname, "-client");
1055 GOTO(out, rc = -ENOMEM);
1057 cfg->cfg_callback = client_lwp_config_process;
1058 cfg->cfg_instance = ll_get_cfg_instance(sb);
1059 rc = lustre_process_log(sb, logname, cfg);
1060 /* need to remove config llog from mgc */
1061 lsi->lsi_lwp_started = 1;
1066 OBD_FREE(logname, MTI_NAME_MAXLEN);
1073 static DEFINE_MUTEX(server_start_lock);
1075 /* Stop MDS/OSS if nobody is using them */
1076 static int server_stop_servers(int lsiflags)
1078 struct obd_device *obd = NULL;
1079 struct obd_type *type = NULL;
1084 mutex_lock(&server_start_lock);
1086 /* Either an MDT or an OST or neither */
1087 /* if this was an MDT, and there are no more MDT's, clean up the MDS */
1088 if (lsiflags & LDD_F_SV_TYPE_MDT) {
1089 obd = class_name2obd(LUSTRE_MDS_OBDNAME);
1090 type = class_search_type(LUSTRE_MDT_NAME);
1091 } else if (lsiflags & LDD_F_SV_TYPE_OST) {
1092 /* if this was an OST, and there are no more OST's, clean up the OSS */
1093 obd = class_name2obd(LUSTRE_OSS_OBDNAME);
1094 type = class_search_type(LUSTRE_OST_NAME);
1097 /* server_stop_servers is a pair of server_start_targets
1098 * Here we put type which was taken at server_start_targets.
1099 * If type is NULL then there is a wrong logic around type or
1100 * type reference. */
1101 LASSERTF(type, "Server flags %d, obd %s\n", lsiflags,
1102 obd ? obd->obd_name : "NULL");
1104 type_last = (type->typ_refcnt == 1);
1106 class_put_type(type);
1107 if (obd != NULL && type_last) {
1109 /* obd_fail doesn't mean much on a server obd */
1110 rc = class_manual_cleanup(obd);
1113 mutex_unlock(&server_start_lock);
1118 int server_mti_print(const char *title, struct mgs_target_info *mti)
1120 PRINT_CMD(PRINT_MASK, "mti %s\n", title);
1121 PRINT_CMD(PRINT_MASK, "server: %s\n", mti->mti_svname);
1122 PRINT_CMD(PRINT_MASK, "fs: %s\n", mti->mti_fsname);
1123 PRINT_CMD(PRINT_MASK, "uuid: %s\n", mti->mti_uuid);
1124 PRINT_CMD(PRINT_MASK, "ver: %d flags: %#x\n",
1125 mti->mti_config_ver, mti->mti_flags);
1129 /* Generate data for registration */
1130 static int server_lsi2mti(struct lustre_sb_info *lsi,
1131 struct mgs_target_info *mti)
1133 struct lnet_process_id id;
1138 if (!IS_SERVER(lsi))
1141 if (strlcpy(mti->mti_svname, lsi->lsi_svname, sizeof(mti->mti_svname))
1142 >= sizeof(mti->mti_svname))
1145 mti->mti_nid_count = 0;
1146 while (LNetGetId(i++, &id) != -ENOENT) {
1147 if (LNET_NETTYP(LNET_NIDNET(id.nid)) == LOLND)
1150 /* server use --servicenode param, only allow specified
1151 * nids be registered */
1152 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) != 0 &&
1153 class_match_nid(lsi->lsi_lmd->lmd_params,
1154 PARAM_FAILNODE, id.nid) < 1)
1157 /* match specified network */
1158 if (!class_match_net(lsi->lsi_lmd->lmd_params,
1159 PARAM_NETWORK, LNET_NIDNET(id.nid)))
1162 mti->mti_nids[mti->mti_nid_count] = id.nid;
1163 mti->mti_nid_count++;
1164 if (mti->mti_nid_count >= MTI_NIDS_MAX) {
1165 CWARN("Only using first %d nids for %s\n",
1166 mti->mti_nid_count, mti->mti_svname);
1171 if (mti->mti_nid_count == 0) {
1172 CERROR("Failed to get NID for server %s, please check whether "
1173 "the target is specifed with improper --servicenode or "
1174 "--network options.\n", mti->mti_svname);
1178 mti->mti_lustre_ver = LUSTRE_VERSION_CODE;
1179 mti->mti_config_ver = 0;
1181 rc = server_name2fsname(lsi->lsi_svname, mti->mti_fsname, NULL);
1185 rc = server_name2index(lsi->lsi_svname, &mti->mti_stripe_index, NULL);
1188 /* Orion requires index to be set */
1189 LASSERT(!(rc & LDD_F_NEED_INDEX));
1190 /* keep only LDD flags */
1191 mti->mti_flags = lsi->lsi_flags & LDD_F_MASK;
1192 if (mti->mti_flags & (LDD_F_WRITECONF | LDD_F_VIRGIN))
1193 mti->mti_flags |= LDD_F_UPDATE;
1194 cplen = strlcpy(mti->mti_params, lsi->lsi_lmd->lmd_params,
1195 sizeof(mti->mti_params));
1196 if (cplen >= sizeof(mti->mti_params))
1201 /* Register an old or new target with the MGS. If needed MGS will construct
1202 startup logs and assign index */
1203 static int server_register_target(struct lustre_sb_info *lsi)
1205 struct obd_device *mgc = lsi->lsi_mgc;
1206 struct mgs_target_info *mti = NULL;
1214 if (!IS_SERVER(lsi))
1221 rc = server_lsi2mti(lsi, mti);
1225 CDEBUG(D_MOUNT, "Registration %s, fs=%s, %s, index=%04x, flags=%#x\n",
1226 mti->mti_svname, mti->mti_fsname,
1227 libcfs_nid2str(mti->mti_nids[0]), mti->mti_stripe_index,
1230 /* if write_conf is true, the registration must succeed */
1231 writeconf = !!(lsi->lsi_flags & (LDD_F_NEED_INDEX | LDD_F_UPDATE));
1232 mti->mti_flags |= LDD_F_OPC_REG;
1235 /* Register the target */
1236 /* FIXME use mgc_process_config instead */
1237 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1238 sizeof(KEY_REGISTER_TARGET),
1239 KEY_REGISTER_TARGET,
1240 sizeof(*mti), mti, NULL);
1242 if (mti->mti_flags & LDD_F_ERROR) {
1243 LCONSOLE_ERROR_MSG(0x160,
1244 "%s: the MGS refuses to allow this server "
1245 "to start: rc = %d. Please see messages on "
1246 "the MGS.\n", lsi->lsi_svname, rc);
1247 } else if (writeconf) {
1248 if ((rc == -ESHUTDOWN || rc == -EIO) && ++tried < 5) {
1249 /* The connection with MGS is not established.
1250 * Try again after 2 seconds. Interruptable. */
1251 set_current_state(TASK_INTERRUPTIBLE);
1252 schedule_timeout(cfs_time_seconds(2));
1253 set_current_state(TASK_RUNNING);
1254 if (!signal_pending(current))
1258 LCONSOLE_ERROR_MSG(0x15f,
1259 "%s: cannot register this server with the MGS: "
1260 "rc = %d. Is the MGS running?\n",
1261 lsi->lsi_svname, rc);
1263 CDEBUG(D_HA, "%s: error registering with the MGS: "
1264 "rc = %d (not fatal)\n", lsi->lsi_svname, rc);
1265 /* reset the error code for non-fatal error. */
1278 * Notify the MGS that this target is ready.
1279 * Used by IR - if the MGS receives this message, it will notify clients.
1281 static int server_notify_target(struct super_block *sb, struct obd_device *obd)
1283 struct lustre_sb_info *lsi = s2lsi(sb);
1284 struct obd_device *mgc = lsi->lsi_mgc;
1285 struct mgs_target_info *mti = NULL;
1291 if (!(IS_SERVER(lsi)))
1297 rc = server_lsi2mti(lsi, mti);
1301 mti->mti_instance = obd->u.obt.obt_instance;
1302 mti->mti_flags |= LDD_F_OPC_READY;
1304 /* FIXME use mgc_process_config instead */
1305 rc = obd_set_info_async(NULL, mgc->u.cli.cl_mgc_mgsexp,
1306 sizeof(KEY_REGISTER_TARGET),
1307 KEY_REGISTER_TARGET,
1308 sizeof(*mti), mti, NULL);
1310 /* Imperative recovery: if the mgs informs us to use IR? */
1311 if (!rc && !(mti->mti_flags & LDD_F_ERROR) &&
1312 (mti->mti_flags & LDD_F_IR_CAPABLE))
1313 lsi->lsi_flags |= LDD_F_IR_CAPABLE;
1322 /** Start server targets: MDTs and OSTs
1324 static int server_start_targets(struct super_block *sb)
1326 struct obd_device *obd;
1327 struct lustre_sb_info *lsi = s2lsi(sb);
1328 struct config_llog_instance cfg;
1329 struct lu_env mgc_env;
1330 struct lu_device *dev;
1331 char *name_service, *obd_name_service = NULL;
1332 struct obd_type *type = NULL;
1336 CDEBUG(D_MOUNT, "starting target %s\n", lsi->lsi_svname);
1338 LASSERTF(IS_MDT(lsi) || IS_OST(lsi), "designed for MDT or OST only\n");
1341 obd_name_service = LUSTRE_MDS_OBDNAME;
1342 name_service = LUSTRE_MDS_NAME;
1344 obd_name_service = LUSTRE_OSS_OBDNAME;
1345 name_service = LUSTRE_OSS_NAME;
1348 /* make sure MDS/OSS is started */
1349 mutex_lock(&server_start_lock);
1350 obd = class_name2obd(obd_name_service);
1352 rc = lustre_start_simple(obd_name_service, name_service,
1354 LUSTRE_MDS_OBDNAME"_uuid" :
1355 LUSTRE_OSS_OBDNAME"_uuid"),
1356 NULL, NULL, NULL, NULL);
1358 mutex_unlock(&server_start_lock);
1359 CERROR("failed to start %s: %d\n",
1360 obd_name_service, rc);
1364 /* hold a type reference and put it at server_stop_servers */
1365 type = class_get_type(IS_MDT(lsi) ?
1366 LUSTRE_MDT_NAME : LUSTRE_OST_NAME);
1367 lsi->lsi_server_started = 1;
1368 mutex_unlock(&server_start_lock);
1369 if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE) &&
1371 OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
1372 msleep(2 * MSEC_PER_SEC);
1375 rc = lu_env_init(&mgc_env, LCT_MG_THREAD);
1377 GOTO(out_stop_service, rc);
1379 /* Set the mgc fs to our server disk. This allows the MGC to
1380 * read and write configs locally, in case it can't talk to the MGS. */
1381 rc = server_mgc_set_fs(&mgc_env, lsi->lsi_mgc, sb);
1385 /* Register with MGS */
1386 rc = server_register_target(lsi);
1390 /* Let the target look up the mount using the target's name
1391 (we can't pass the sb or mnt through class_process_config.) */
1392 rc = server_register_mount(lsi->lsi_svname, sb);
1396 /* Start targets using the llog named for the target */
1397 memset(&cfg, 0, sizeof(cfg));
1398 cfg.cfg_callback = class_config_llog_handler;
1399 cfg.cfg_sub_clds = CONFIG_SUB_SERVER;
1400 rc = lustre_process_log(sb, lsi->lsi_svname, &cfg);
1402 CERROR("failed to start server %s: %d\n",
1403 lsi->lsi_svname, rc);
1404 /* Do NOT call server_deregister_mount() here. This makes it
1405 * impossible to find mount later in cleanup time and leaves
1406 * @lsi and othder stuff leaked. -umka */
1410 obd = class_name2obd(lsi->lsi_svname);
1412 CERROR("no server named %s was started\n", lsi->lsi_svname);
1413 GOTO(out_mgc, rc = -ENXIO);
1416 if (IS_OST(lsi) || IS_MDT(lsi)) {
1417 rc = lustre_start_lwp(sb);
1419 CERROR("%s: failed to start LWP: %d\n",
1420 lsi->lsi_svname, rc);
1425 server_notify_target(sb, obd);
1427 /* calculate recovery timeout, do it after lustre_process_log */
1428 server_calc_timeout(lsi, obd);
1430 /* log has been fully processed, let clients connect */
1431 dev = obd->obd_lu_dev;
1432 if (dev && dev->ld_ops->ldo_prepare) {
1435 rc = lu_env_init(&env, dev->ld_type->ldt_ctx_tags);
1437 struct lu_context session_ctx;
1439 lu_context_init(&session_ctx, LCT_SERVER_SESSION);
1440 session_ctx.lc_thread = NULL;
1441 lu_context_enter(&session_ctx);
1442 env.le_ses = &session_ctx;
1444 rc = dev->ld_ops->ldo_prepare(&env, NULL, dev);
1447 lu_context_exit(&session_ctx);
1448 lu_context_fini(&session_ctx);
1452 /* abort recovery only on the complete stack:
1453 * many devices can be involved */
1454 if ((lsi->lsi_lmd->lmd_flags & LMD_FLG_ABORT_RECOV) &&
1455 (OBP(obd, iocontrol))) {
1456 obd_iocontrol(OBD_IOC_ABORT_RECOVERY, obd->obd_self_export, 0,
1461 /* Release the mgc fs for others to use */
1462 server_mgc_clear_fs(&mgc_env, lsi->lsi_mgc);
1464 lu_env_fini(&mgc_env);
1466 /* in case of error upper function call
1467 * server_put_super->server_stop_servers()
1473 static int lsi_prepare(struct lustre_sb_info *lsi)
1475 const char *osd_type;
1482 LASSERT(lsi->lsi_lmd);
1484 /* The server name is given as a mount line option */
1485 if (lsi->lsi_lmd->lmd_profile == NULL) {
1486 LCONSOLE_ERROR("Can't determine server name\n");
1490 /* Determine osd type */
1491 if (lsi->lsi_lmd->lmd_osd_type == NULL) {
1492 osd_type = LUSTRE_OSD_LDISKFS_NAME;
1495 osd_type = lsi->lsi_lmd->lmd_osd_type;
1496 fstype = lsi->lsi_lmd->lmd_osd_type;
1499 if (strlen(lsi->lsi_lmd->lmd_profile) >= sizeof(lsi->lsi_svname) ||
1500 strlen(osd_type) >= sizeof(lsi->lsi_osd_type) ||
1501 strlen(fstype) >= sizeof(lsi->lsi_fstype))
1502 RETURN(-ENAMETOOLONG);
1504 strlcpy(lsi->lsi_svname, lsi->lsi_lmd->lmd_profile,
1505 sizeof(lsi->lsi_svname));
1506 strlcpy(lsi->lsi_osd_type, osd_type, sizeof(lsi->lsi_osd_type));
1507 /* XXX: a temp. solution for components using ldiskfs
1508 * to be removed in one of the subsequent patches */
1509 strlcpy(lsi->lsi_fstype, fstype, sizeof(lsi->lsi_fstype));
1511 /* Determine server type */
1512 rc = server_name2index(lsi->lsi_svname, &index, NULL);
1514 if (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) {
1515 /* Assume we're a bare MGS */
1517 lsi->lsi_lmd->lmd_flags |= LMD_FLG_NOSVC;
1519 LCONSOLE_ERROR("Can't determine server type of '%s'\n",
1524 lsi->lsi_flags |= rc;
1526 /* Add mount line flags that used to be in ldd:
1527 * writeconf, mgs, anything else?
1529 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_WRITECONF) ?
1530 LDD_F_WRITECONF : 0;
1531 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_VIRGIN) ?
1533 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_UPDATE) ?
1535 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_MGS) ?
1536 LDD_F_SV_TYPE_MGS : 0;
1537 lsi->lsi_flags |= (lsi->lsi_lmd->lmd_flags & LMD_FLG_NO_PRIMNODE) ?
1538 LDD_F_NO_PRIMNODE : 0;
1543 /*************** server mount ******************/
1545 /** Start the shutdown of servers at umount.
1547 static void server_put_super(struct super_block *sb)
1549 struct lustre_sb_info *lsi = s2lsi(sb);
1550 struct obd_device *obd;
1551 char *tmpname, *extraname = NULL;
1553 int lsiflags = lsi->lsi_flags;
1554 bool stop_servers = lsi->lsi_server_started;
1557 LASSERT(IS_SERVER(lsi));
1559 tmpname_sz = strlen(lsi->lsi_svname) + 1;
1560 OBD_ALLOC(tmpname, tmpname_sz);
1561 memcpy(tmpname, lsi->lsi_svname, tmpname_sz);
1562 CDEBUG(D_MOUNT, "server put_super %s\n", tmpname);
1563 if (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC))
1564 snprintf(tmpname, tmpname_sz, "MGS");
1566 /* disconnect the lwp first to drain off the inflight request */
1567 if (IS_OST(lsi) || IS_MDT(lsi)) {
1570 rc = lustre_disconnect_lwp(sb);
1571 if (rc != 0 && rc != -ETIMEDOUT &&
1572 rc != -ENOTCONN && rc != -ESHUTDOWN)
1573 CWARN("%s: failed to disconnect lwp: rc= %d\n",
1577 /* Stop the target */
1578 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
1579 (IS_MDT(lsi) || IS_OST(lsi))) {
1580 struct lustre_profile *lprof = NULL;
1582 /* tell the mgc to drop the config log */
1583 lustre_end_log(sb, lsi->lsi_svname, NULL);
1585 /* COMPAT_146 - profile may get deleted in mgc_cleanup.
1586 If there are any setup/cleanup errors, save the lov
1587 name for safety cleanup later. */
1588 lprof = class_get_profile(lsi->lsi_svname);
1589 if (lprof != NULL) {
1590 if (lprof->lp_dt != NULL) {
1591 OBD_ALLOC(extraname, strlen(lprof->lp_dt) + 1);
1592 strncpy(extraname, lprof->lp_dt,
1593 strlen(lprof->lp_dt) + 1);
1595 class_put_profile(lprof);
1598 obd = class_name2obd(lsi->lsi_svname);
1600 CDEBUG(D_MOUNT, "stopping %s\n", obd->obd_name);
1601 if (lsiflags & LSI_UMOUNT_FAILOVER)
1603 /* We can't seem to give an error return code
1604 * to .put_super, so we better make sure we clean up! */
1606 class_manual_cleanup(obd);
1607 if (OBD_FAIL_PRECHECK(OBD_FAIL_OBD_STOP_MDS_RACE)) {
1609 server_name2index(lsi->lsi_svname, &idx, NULL);
1610 /* sleeping for MDT0001 */
1612 OBD_RACE(OBD_FAIL_OBD_STOP_MDS_RACE);
1615 CERROR("no obd %s\n", lsi->lsi_svname);
1616 server_deregister_mount(lsi->lsi_svname);
1620 /* If they wanted the mgs to stop separately from the mdt, they
1621 should have put it on a different device. */
1623 /* if MDS start with --nomgs, don't stop MGS then */
1624 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS))
1625 server_stop_mgs(sb);
1628 if (IS_OST(lsi) || IS_MDT(lsi)) {
1629 if (lustre_stop_lwp(sb) < 0)
1630 CERROR("%s: failed to stop lwp!\n", tmpname);
1633 /* Clean the mgc and sb */
1634 lustre_common_put_super(sb);
1636 /* wait till all in-progress cleanups are done
1637 * specifically we're interested in ofd cleanup
1639 obd_zombie_barrier();
1641 /* Stop the servers (MDS, OSS) if no longer needed. We must wait
1642 until the target is really gone so that our type refcount check
1645 server_stop_servers(lsiflags);
1647 /* In case of startup or cleanup err, stop related obds */
1649 obd = class_name2obd(extraname);
1651 CWARN("Cleaning orphaned obd %s\n", extraname);
1653 class_manual_cleanup(obd);
1655 OBD_FREE(extraname, strlen(extraname) + 1);
1658 LCONSOLE_WARN("server umount %s complete\n", tmpname);
1659 OBD_FREE(tmpname, tmpname_sz);
1663 /** Called only for 'umount -f'
1665 static void server_umount_begin(struct super_block *sb)
1667 struct lustre_sb_info *lsi = s2lsi(sb);
1670 CDEBUG(D_MOUNT, "umount -f\n");
1671 /* umount = failover
1673 no third way to do non-force, non-failover */
1674 lsi->lsi_flags &= ~LSI_UMOUNT_FAILOVER;
1678 static int server_statfs(struct dentry *dentry, struct kstatfs *buf)
1680 struct super_block *sb = dentry->d_sb;
1681 struct lustre_sb_info *lsi = s2lsi(sb);
1682 struct obd_statfs statfs;
1686 if (lsi->lsi_dt_dev) {
1687 rc = dt_statfs(NULL, lsi->lsi_dt_dev, &statfs);
1689 statfs_unpack(buf, &statfs);
1690 buf->f_type = sb->s_magic;
1696 buf->f_type = sb->s_magic;
1697 buf->f_bsize = sb->s_blocksize;
1703 buf->f_namelen = NAME_MAX;
1707 #ifdef HAVE_SUPEROPS_USE_DENTRY
1708 int server_show_options(struct seq_file *seq, struct dentry *dentry)
1710 int server_show_options(struct seq_file *seq, struct vfsmount *vfs)
1713 struct lustre_sb_info *lsi;
1714 struct lustre_mount_data *lmd;
1716 #ifdef HAVE_SUPEROPS_USE_DENTRY
1717 LASSERT(seq != NULL && dentry != NULL);
1718 lsi = s2lsi(dentry->d_sb);
1720 LASSERT(seq != NULL && vfs != NULL);
1721 lsi = s2lsi(vfs->mnt_sb);
1725 seq_printf(seq, ",svname=%s", lmd->lmd_profile);
1727 if (lmd->lmd_flags & LMD_FLG_ABORT_RECOV)
1728 seq_puts(seq, ",abort_recov");
1730 if (lmd->lmd_flags & LMD_FLG_NOIR)
1731 seq_puts(seq, ",noir");
1733 if (lmd->lmd_flags & LMD_FLG_NOSVC)
1734 seq_puts(seq, ",nosvc");
1736 if (lmd->lmd_flags & LMD_FLG_NOMGS)
1737 seq_puts(seq, ",nomgs");
1739 if (lmd->lmd_flags & LMD_FLG_NOSCRUB)
1740 seq_puts(seq, ",noscrub");
1741 if (lmd->lmd_flags & LMD_FLG_SKIP_LFSCK)
1742 seq_puts(seq, ",skip_lfsck");
1744 if (lmd->lmd_flags & LMD_FLG_DEV_RDONLY)
1745 seq_puts(seq, ",rdonly_dev");
1747 if (lmd->lmd_flags & LMD_FLG_MGS)
1748 seq_puts(seq, ",mgs");
1750 if (lmd->lmd_mgs != NULL)
1751 seq_printf(seq, ",mgsnode=%s", lmd->lmd_mgs);
1753 if (lmd->lmd_osd_type != NULL)
1754 seq_printf(seq, ",osd=%s", lmd->lmd_osd_type);
1756 if (lmd->lmd_opts != NULL) {
1758 seq_puts(seq, lmd->lmd_opts);
1764 /** The operations we support directly on the superblock:
1765 * mount, umount, and df.
1767 static struct super_operations server_ops = {
1768 .put_super = server_put_super,
1769 .umount_begin = server_umount_begin, /* umount -f */
1770 .statfs = server_statfs,
1771 .show_options = server_show_options,
1775 * Xattr support for Lustre servers
1777 #ifdef HAVE_IOP_XATTR
1778 static ssize_t lustre_getxattr(struct dentry *dentry, const char *name,
1779 void *buffer, size_t size)
1781 if (!selinux_is_enabled())
1786 static int lustre_setxattr(struct dentry *dentry, const char *name,
1787 const void *value, size_t size, int flags)
1793 static ssize_t lustre_listxattr(struct dentry *d_entry, char *name,
1799 static bool is_cmd_supported(unsigned int command)
1811 static long server_ioctl(struct file *filp, unsigned int command,
1814 struct file active_filp;
1815 struct inode *inode = file_inode(filp);
1816 struct lustre_sb_info *lsi = s2lsi(inode->i_sb);
1817 struct super_block *dd_sb = dt_mnt_sb_get(lsi->lsi_dt_dev);
1818 struct inode *active_inode;
1819 int err = -EOPNOTSUPP;
1821 if (IS_ERR(dd_sb) || !is_cmd_supported(command))
1824 active_inode = igrab(dd_sb->s_root->d_inode);
1828 active_filp.f_inode = active_inode;
1829 if (active_inode->i_fop && active_inode->i_fop->unlocked_ioctl)
1830 err = active_inode->i_fop->unlocked_ioctl(&active_filp,
1836 static const struct inode_operations server_inode_operations = {
1837 #ifdef HAVE_IOP_XATTR
1838 .setxattr = lustre_setxattr,
1839 .getxattr = lustre_getxattr,
1841 .listxattr = lustre_listxattr,
1844 static const struct file_operations server_file_operations = {
1845 .unlocked_ioctl = server_ioctl,
1848 #define log2(n) ffz(~(n))
1849 #define LUSTRE_SUPER_MAGIC 0x0BD00BD1
1851 static int server_fill_super_common(struct super_block *sb)
1853 struct inode *root = NULL;
1856 CDEBUG(D_MOUNT, "Server sb, dev=%d\n", (int)sb->s_dev);
1858 sb->s_blocksize = 4096;
1859 sb->s_blocksize_bits = log2(sb->s_blocksize);
1860 sb->s_magic = LUSTRE_SUPER_MAGIC;
1861 sb->s_maxbytes = 0; /* we don't allow file IO on server mountpoints */
1862 sb->s_flags |= SB_RDONLY;
1863 sb->s_op = &server_ops;
1865 root = new_inode(sb);
1867 CERROR("Can't make root inode\n");
1871 /* returns -EIO for every operation */
1872 /* make_bad_inode(root); -- badness - can't umount */
1873 /* apparently we need to be a directory for the mount to finish */
1874 root->i_mode = S_IFDIR;
1875 root->i_op = &server_inode_operations;
1876 root->i_fop = &server_file_operations;
1877 sb->s_root = d_make_root(root);
1879 CERROR("%s: can't make root dentry\n", sb->s_id);
1886 static int osd_start(struct lustre_sb_info *lsi, unsigned long mflags)
1888 struct lustre_mount_data *lmd = lsi->lsi_lmd;
1889 struct obd_device *obd;
1890 struct dt_device_param p;
1891 char flagstr[20 + 1 + 10 + 1];
1896 "Attempting to start %s, type=%s, lsifl=%x, mountfl=%lx\n",
1897 lsi->lsi_svname, lsi->lsi_osd_type, lsi->lsi_flags, mflags);
1899 sprintf(lsi->lsi_osd_obdname, "%s-osd", lsi->lsi_svname);
1900 strcpy(lsi->lsi_osd_uuid, lsi->lsi_osd_obdname);
1901 strcat(lsi->lsi_osd_uuid, "_UUID");
1902 snprintf(flagstr, sizeof(flagstr), "%lu:%u", mflags, lmd->lmd_flags);
1904 obd = class_name2obd(lsi->lsi_osd_obdname);
1906 rc = lustre_start_simple(lsi->lsi_osd_obdname,
1908 lsi->lsi_osd_uuid, lmd->lmd_dev,
1909 flagstr, lsi->lsi_lmd->lmd_opts,
1913 obd = class_name2obd(lsi->lsi_osd_obdname);
1916 CDEBUG(D_MOUNT, "%s already started\n", lsi->lsi_osd_obdname);
1917 /* but continue setup to allow special case of MDT and internal
1918 * MGT being started separately. */
1919 if (!((IS_MGS(lsi) && (lsi->lsi_lmd->lmd_flags &
1921 (IS_MDT(lsi) && (lsi->lsi_lmd->lmd_flags &
1926 rc = obd_connect(NULL, &lsi->lsi_osd_exp,
1927 obd, &obd->obd_uuid, NULL, NULL);
1931 class_manual_cleanup(obd);
1932 lsi->lsi_dt_dev = NULL;
1936 LASSERT(obd->obd_lu_dev);
1937 lu_device_get(obd->obd_lu_dev);
1938 lsi->lsi_dt_dev = lu2dt_dev(obd->obd_lu_dev);
1939 LASSERT(lsi->lsi_dt_dev);
1941 /* set disk context for llog usage */
1942 OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1943 obd->obd_lvfs_ctxt.dt = lsi->lsi_dt_dev;
1945 dt_conf_get(NULL, lsi->lsi_dt_dev, &p);
1950 /** Fill in the superblock info for a Lustre server.
1951 * Mount the device with the correct options.
1952 * Read the on-disk config file.
1953 * Start the services.
1955 int server_fill_super(struct super_block *sb)
1957 struct lustre_sb_info *lsi = s2lsi(sb);
1961 /* to simulate target mount race */
1962 OBD_RACE(OBD_FAIL_TGT_MOUNT_RACE);
1964 rc = lsi_prepare(lsi);
1970 /* Start low level OSD */
1971 rc = osd_start(lsi, sb->s_flags);
1973 CERROR("Unable to start osd on %s: %d\n",
1974 lsi->lsi_lmd->lmd_dev, rc);
1979 CDEBUG(D_MOUNT, "Found service %s on device %s\n",
1980 lsi->lsi_svname, lsi->lsi_lmd->lmd_dev);
1982 if (class_name2obd(lsi->lsi_svname)) {
1983 LCONSOLE_ERROR_MSG(0x161, "The target named %s is already "
1984 "running. Double-mount may have compromised"
1985 " the disk journal.\n",
1991 /* Start MGS before MGC */
1992 if (IS_MGS(lsi) && !(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOMGS)) {
1993 rc = server_start_mgs(sb);
1998 /* Start MGC before servers */
1999 rc = lustre_start_mgc(sb);
2003 /* Set up all obd devices for service */
2004 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOSVC) &&
2005 (IS_OST(lsi) || IS_MDT(lsi))) {
2006 rc = server_start_targets(sb);
2008 CERROR("Unable to start targets: %d\n", rc);
2011 /* FIXME overmount client here, or can we just start a
2012 * client log and client_fill_super on this sb? We
2013 * need to make sure server_put_super gets called too
2014 * - ll_put_super calls lustre_common_put_super; check
2015 * there for LSI_SERVER flag, call s_p_s if so.
2017 * Probably should start client from new thread so we
2018 * can return. Client will not finish until all
2019 * servers are connected. Note - MGS-only server does
2020 * NOT get a client, since there is no lustre fs
2021 * associated - the MGS is for all lustre fs's */
2024 rc = server_fill_super_common(sb);
2030 /* We jump here in case of failure while starting targets or MGS.
2031 * In this case we can't just put @mnt and have to do real cleanup
2032 * with stoping targets, etc. */
2033 server_put_super(sb);
2038 * Calculate timeout value for a target.
2040 void server_calc_timeout(struct lustre_sb_info *lsi, struct obd_device *obd)
2042 struct lustre_mount_data *lmd;
2046 bool has_ir = !!(lsi->lsi_flags & LDD_F_IR_CAPABLE);
2047 int min = OBD_RECOVERY_TIME_MIN;
2049 LASSERT(IS_SERVER(lsi));
2053 soft = lmd->lmd_recovery_time_soft;
2054 hard = lmd->lmd_recovery_time_hard;
2055 has_ir = has_ir && !(lmd->lmd_flags & LMD_FLG_NOIR);
2056 obd->obd_no_ir = !has_ir;
2060 soft = OBD_RECOVERY_TIME_SOFT;
2062 hard = OBD_RECOVERY_TIME_HARD;
2064 /* target may have ir_factor configured. */
2065 factor = OBD_IR_FACTOR_DEFAULT;
2066 if (obd->obd_recovery_ir_factor)
2067 factor = obd->obd_recovery_ir_factor;
2070 int new_soft = soft;
2072 /* adjust timeout value by imperative recovery */
2073 new_soft = (soft * factor) / OBD_IR_FACTOR_MAX;
2074 /* make sure the timeout is not too short */
2075 new_soft = max(min, new_soft);
2077 LCONSOLE_INFO("%s: Imperative Recovery enabled, recovery "
2078 "window shrunk from %d-%d down to %d-%d\n",
2079 obd->obd_name, soft, hard, new_soft, hard);
2083 LCONSOLE_INFO("%s: Imperative Recovery not enabled, recovery "
2084 "window %d-%d\n", obd->obd_name, soft, hard);
2088 obd->obd_recovery_timeout = max_t(time64_t, obd->obd_recovery_timeout,
2090 obd->obd_recovery_time_hard = hard;
2091 obd->obd_recovery_ir_factor = factor;